On-line sound event detection and recognition based on adaptive background model for robot audition

It´s a natural and convenient way for a robot to interact with outside by robot´s ears (i.e. microphones) based on correctly detection and recognition of a sound event. This paper considers sound event detection and recognition in indoor environment where there are varying noises around a robot. To handle the problem of varying background noises, a novel sound event detection and recognition system is developed. Background model update and re-estimation methods are respectively proposed to handle the situations when background noises change slightly or completely. Recognition is then conducted based on the detected sound event by matching it with the noise-corrupted models generated by our proposed combining method modified Parallel Model Combination method (mPMC). mPMC allows modeling the background noise by Gaussian Mixture Model (GMM) of multiple components and can represent the background noise more precisely compared to Single Gaussian Model (SGM). Experimental results show that our adaptive background modeling method attains excellent detection performance in noise-varying conditions and the recognition performance of our proposed mPMC using GMM also outperforms the conventional PMC using SGM in real-world environment with noise varying.